Brain-SAD: A Brain-Inspired Safe Autonomous Driving Control Framework with Dynamic Fear-Oriented Constraint on Dual-Policy

Autonomous-driving policies must balance progress with safety as traffic conditions change. This paper proposes Brain-SAD, a reinforcement-learning framework that adapts its safety constraints to the current interaction scene rather than applying the same cost rules or action boundaries everywhere. Its central idea is to estimate a dynamic “fear” signal from predicted interactions with nearby vehicles, then use that signal to choose between a long-term policy for ordinary driving and a short-term policy for urgent collision defense. The authors report that Brain-SAD improves success rates, task-completion time, and collision-recovery time in their experiments, while remaining more reliable across continuous intersections of changing complexity. Read the full paper on Brain-SAD.

Why static safety constraints can fail

Constrained reinforcement learning generally seeks to maximize driving reward while keeping risk below a specified limit. The paper describes two common approaches. Soft-constrained, or primal-dual, methods limit cumulative cost over a trajectory. A policy may be rewarded for progress but penalized when its actions contribute to risk. Hard-constrained methods instead enforce safety at each step, often by projecting a risky action into a precomputed feasible region.
Both approaches can be useful, but the authors argue that their constraints are often insufficiently adaptive. In soft-constrained systems, cost may be assigned using a static mapping from the current state to a risk value. In hard-constrained systems, the safe-region boundary may be estimated offline and then kept fixed. As a result, similar actions can receive similar treatment even when the surrounding vehicles have different intentions.
The paper illustrates this problem with two interaction scenes. In one, the ego vehicle can continue straight safely. In another, a neighboring vehicle has an intention to turn, making the same behavior dangerous. A static state-level cost or fixed projection boundary may not adequately reflect that difference. The authors therefore focus on constraints that respond to both the current scene and the likely consequences of an action.

How the dual-policy framework works

Brain-SAD first builds an interaction-aware ego state. For each neighboring vehicle, it records relative distance, bearing angle, relative velocity, and relative heading angle. These intent-related vectors are combined with the ego vehicle’s own velocity and heading, giving the controller information about both the vehicle it controls and the surrounding traffic.
The system then imagines future interaction trajectories. It samples an action from the current long-term policy and uses an ensemble of structurally identical but parameter-independent multilayer perceptrons as environment models. These models roll the scene forward over multiple future steps. The imagined trajectories are used to estimate the likely interaction risk associated with the current situation and candidate action.
This estimate is converted into a dynamic fear reaction, inspired by the paper’s analogy to the amygdala and its communication with the prefrontal cortex. The fear signal is not merely a fixed label attached to the current state. It is intended to capture how the interaction may evolve after an action is taken. A learned policy-selection mechanism uses the signal to determine whether the vehicle should remain with the regular long-term controller or switch to an emergency-oriented short-term controller.

Two forms of dynamic constraint

When the fear reaction is mild, Brain-SAD uses its long-term policy for regular interaction. This policy applies a dynamic action fear-cost constraint. The cost is coupled to the predicted impact of actions under the current interaction scenario, rather than being determined only by a static state-to-cost rule. The authors describe this as an online correction of the action distribution: the policy can continue optimizing for longer-term driving objectives while adapting its safety pressure to changing traffic.
When the fear reaction is strong, the framework activates a short-term policy for urgent collision defense, including emergency braking. This policy constructs a dynamic fear boundary for the feasible action region from risky neighboring vehicles. Instead of relying on one fixed projection boundary, it can tighten or loosen the set of permitted actions according to the detected danger. The goal is to select a safe action directly and efficiently when there is insufficient time for ordinary long-horizon optimization.
The brain-inspired terminology maps these components onto different action-control pathways. The long-term controller is associated with the direct and indirect striatal pathways, while the emergency controller is associated with the hyper-direct pathway through the subthalamic nucleus. The authors present this as a functional inspiration for organizing policy selection and constraint handling, rather than as a claim that the artificial system reproduces the biological brain.

What the reported results show

According to the paper, Brain-SAD outperforms existing methods in the evaluated autonomous-driving settings. It achieves a higher success rate while requiring less time for task completion and collision recovery. The framework also shows stronger reliability across continuous intersections whose complexity fluctuates, suggesting that its dynamic constraints help maintain more stable action behavior when the interaction context changes.
The reported advantage is attributed to the collaboration between the two policies. The long-term policy handles ordinary scene evolution, while the short-term policy responds to severe and immediate collision risk. The fear-oriented signal provides the link between perception, policy selection, and constraint adaptation.
The supplied paper material does not include the numerical values of the experiments, the complete training procedure, or detailed information about the benchmark configurations. It therefore supports the authors’ qualitative claims about comparative performance and reliability, but not a more precise assessment of effect size or generalization beyond the tested scenarios. The main contribution is the framework’s design: safety constraints are treated as online, scenario-dependent quantities tied to predicted action impact rather than as static costs or fixed feasible-region boundaries.